Gemma 4 models

44 Gemma 4 models from Google on Hugging Face, from 420M to 32.7B parameters, published by Google in BF16, with contexts from 128K to 256K tokens. The smallest official model, gemma-4-26B-A4B-it-assistant, needs about 0.9 GB of VRAM at its published precision; the cheapest live fit is RTX 3070 at $0.088/hr.

Part of the Gemma series ยท Previous generation: Gemma 3n

Pick a size

One row per official Gemma 4 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.

ModelParametersNative VRAMFP8 VRAMINT4 VRAMLive GPU fit (native)Est. $/hrKV cache at 32K
gemma-4-26B-A4B-it-assistant420M0.9 GB0.5 GB0.2 GBRTX 3070$0.088/hr0.27 GB
gemma-4-31B-it-assistant470M1.0 GB0.5 GB0.3 GBRTX 3070$0.088/hr0.55 GB
gemma-4-E2B-it5.1B11.5 GB5.7 GB2.9 GBRTX A4000$0.137/hr0.23 GB
gemma-4-E4B-it8.0B17.9 GB8.9 GB4.5 GBRTX A5000$0.176/hr0.47 GB
gemma-4-12B-it12.0B26.7 GB13.4 GB6.7 GBRTX A6000$0.363/hr2.3 GB
gemma-4-26B-A4B-it25.8B57.7 GB28.8 GB14.4 GBA100$1.31/hr1.4 GB
gemma-4-26B-A4B-it-qat-q4_0-unquantized26.5B59.3 GB29.7 GB14.8 GBA100$1.31/hr1.4 GB
gemma-4-31B-it31.3B69.9 GB35.0 GB17.5 GBA100$1.31/hr5.8 GB
gemma-4-31B32.7B73.1 GB36.5 GB18.3 GBA100$1.31/hr5.8 GB

VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.

Official models (15)

ModelParametersPublished asVRAM neededLive GPU fitEst. $/hr
gemma-4-26B-A4B-it-assistant420MBF160.9 GBRTX 3070$0.088/hr
gemma-4-12B-it-assistant423MBF160.9 GBRTX 3070$0.088/hr
gemma-4-12B-it-qat-q4_0-unquantized-assistant423MBF160.9 GBRTX 3070$0.088/hr
gemma-4-31B-it-assistant470MBF161.0 GBRTX 3070$0.088/hr
gemma-4-31B-it-qat-q4_0-unquantized-assistant470MBF161.0 GBRTX 3070$0.088/hr
gemma-4-E2B-it5.1BBF1611.5 GBRTX A4000$0.137/hr
gemma-4-E4B8.0BBF1617.9 GBRTX A5000$0.176/hr
gemma-4-E4B-it8.0BBF1617.9 GBRTX A5000$0.176/hr
gemma-4-12B12.0BBF1626.7 GBRTX A6000$0.363/hr
gemma-4-12B-it12.0BBF1626.7 GBRTX A6000$0.363/hr
gemma-4-12B-it-qat-q4_0-unquantized12.0BBF1626.7 GBRTX A6000$0.363/hr
gemma-4-26B-A4B-it25.8BBF1657.7 GBA100$1.31/hr
gemma-4-26B-A4B-it-qat-q4_0-unquantized26.5BBF1659.3 GBA100$1.31/hr
gemma-4-31B-it31.3BBF1669.9 GBA100$1.31/hr
gemma-4-31B32.7BBF1673.1 GBA100$1.31/hr

Quantized versions (13)

ModelParametersPublished asVRAM neededLive GPU fitEst. $/hr
gemma-4-E4B-it-W4A16-AutoRound-GPTQ4.0BBF168.9 GBRTX A4000$0.137/hr
gemma-4-E2B-it-unsloth-bnb-4bit5.1BBF1611.5 GBRTX A4000$0.137/hr
gemma-4-E2B-it-qat-w4a16-ct5.6BBF1612.4 GBRTX A4000$0.137/hr
Gemma-4-E4B-DECKARD-HERETIC-NVFP46.2BBF1613.9 GBRTX A4000$0.137/hr
gemma-4-E4B-it-unsloth-bnb-4bit8.0BBF1617.9 GBRTX A5000$0.176/hr
gemma-4-E4B-it-AWQ-INT48.2BF1618.3 GBRTX A5000$0.176/hr
gemma-4-E4B-it-qat-w4a16-ct8.7BBF1619.5 GBRTX A5000$0.176/hr
gemma-4-12B-it-FP8-Dynamic13.0BF8_E4M314.5 GBRTX 5060 Ti$0.180/hr
Gemma-4-31B-IT-NVFP420.9BBF1646.6 GBRTX A6000$0.363/hr
rune-26b-a4b-GGUF25.8BBF1657.7 GBA100$1.31/hr
gemma-4-26B-A4B-it-FP8-dynamic26.5BF8_E4M329.7 GBRTX 5000 Ada$0.539/hr
gemma-4-31B-it-FP8-block31.3BF8_E4M335.0 GBL40$0.759/hr
gemma-4-31B-it-FP8-dynamic31.3BF8_E4M335.0 GBL40$0.759/hr

Fine-tunes and community models (16)

ModelParametersPublished asVRAM neededLive GPU fitEst. $/hr
gemma-4-26B-A4B-it-DFlash430MBF161.0 GBRTX 3070$0.088/hr
gemma-4-31B-it-DFlash1.5BBF163.4 GBRTX 3070$0.088/hr
gemma4-e4b-claims-comparison7.9BBF1617.7 GBRTX A5000$0.176/hr
gemma-4-E4B-it-OBLITERATED8.0BF1617.9 GBRTX A5000$0.176/hr
gemma-4-E4B-it-uncensored-heretic8.0BBF1617.9 GBRTX A5000$0.176/hr
Jev-Omni12.0BBF1626.7 GBRTX A6000$0.363/hr
humanizer12.0BBF1626.7 GBRTX A6000$0.363/hr
Gemma-4-12B-OBLITERATED12.0BBF1626.7 GBRTX A6000$0.363/hr
gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau212.0BF1626.7 GBV100$0.220/hr
GEV-26B-Decide25.8BBF1657.7 GBA100$1.31/hr
Orion-26B-A4B-v125.8BBF1657.7 GBA100$1.31/hr
Orion-26B-A4B-v1.125.8BBF1657.7 GBA100$1.31/hr
SOV-Transcribe-126.6BBF1659.5 GBA100$1.31/hr
Artemis-31B-v1.231.3BBF1669.9 GBA100$1.31/hr
OTel-2.0-LLM-31B-IT32.1BBF1671.8 GBA100$1.31/hr
gemma-4-31B-it-uncensored32.7BBF1673.1 GBA100$1.31/hr

VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.

Run and fine-tune Gemma 4

More from Google

Start ComfyUI on a rented GPU.

ยฉ 2026 Aquanode. All rights reserved.

All trademarks, logos and brand names are the property of their respective owners.